Study-isolated evaluation of literature-derived UHPC strength prediction — data and code

Description

This directory contains the data, model artefacts, source code and documentation for a study on machine-learning prediction of ultra-high performance concrete (UHPC) properties and on the inverse design of UHPC mixtures. The central question of the study is whether literature-derived UHPC strength models remain accurate when the evaluation protocol removes information leakage between records that originate from the same publication.

Steps to reproduce

1. Layout. The files are distributed flat. Restore the assembled directory layout with the snippet in FILE_INDEX.md before running anything; the paths below use that layout. 2. Environment (docs@@reproducibility.md, section 1). Windows 11; Python 3.14.2; numpy 2.4.2, pandas 3.0.1, scikit-learn 1.9.0, xgboost 3.4.1, catboost 1.2.10, optuna 4.9.0, joblib 1.5.3, SHAP, matplotlib. Global seed 20260827. The frozen run used an NVIDIA RTX 4080 Laptop GPU (driver 616.56). 3. Verify the download. Run the verification snippet in FILE_INDEX.md and expect "checked 281 files, 0 mismatches". The snapshot digest is 9a0045feb1304b026c8f3dc95256041897a217ece7a060a03054249bd450a7cd and the source matrix digest is 8e26c9427ee31acdb7d1fe7caa6172cb115c331684ca301ca3f8fe33a0166b00. 4. Structural check. Read data/UHPC_dataset_snapshot.csv with pandas: 2029 rows, 132 columns, 170 canonical_paper_group_id groups, record_id unique. 5. Regenerate the reported results, in this order: python code\final_pipeline\step4_split_manifest.py python code\final_pipeline\step2_run_main_matrix.py python code\final_pipeline\step6_eval_stability.py python code\final_pipeline\step7_deploy_pack.py python code\final_pipeline\step8_candidates.py python code\final_pipeline\step9_reports.py python code\final_pipeline\step9b_model_card.py python code\final_pipeline\step3_old_pipeline_repro.py python code\cbm_reinforcement\src\run_all.py python -m pytest code\cbm_reinforcement\src\test_cbm_reinforcement.py -q python code\figures\redraw_validated_figures.py python code\figures\draw_M_extension_figures.py 6. Check the headline values against reports@@stability_significance_summary.json: best-cell MAE 16.8556 MPa, group-macro MAE 17.6155 MPa, RMSE 22.0316 MPa, R2 0.4897, 90% interval coverage 0.9397, pooled MAE 17.9002 MPa on 1344 records under the strict supergroup resplit. 7. To reanalyse without retraining, read the frozen tables under predictions@@; re-running the estimators on other hardware may not reproduce predictions bitwise. 8. Optional audit of the dataset derivation: set SRC_V6 in code\final_pipeline\step1_build_snapshot.py to data/UHPC_dataset_v6_source.csv and run it; a correct rebuild reproduces the snapshot digest.

Institutions

• Xi'an University of Architecture and Technology

  Shaanxi, Xi'an

Categories

Funders

• National Natural Science Foundation of China

  Beijing, Beijing

  Grant ID: Award No. 52378190
• Xi'an University of Architecture and Technology

  Shaanxi, Xi'an

  Grant ID: Entrepreneurship Key Cultivation Program of Xi'an University of Architecture and Technology，Program No. 26CY013

Related Links

Licence